About the VAE mel-spectrogram model

[wblgers]

Hi,

How many hours of training data you prepare to train the VAE mel-spectrogram?
And is this model to convert mel-spectrogram into tokens? And the acoustic model predicts tokens instead of mel-spectrogram?

Thanks!

[leng-yue]

16kh audios probably.